Cohere Skipped Consumer Chatbots and AGI, Then Hit $240M ARR: Nick Frosst's Enterprise AI Story

October 5, 2026

TL;DR: Cohere is a Toronto-based company that trains large language models and sells them to businesses, not consumers. It was founded in 2019 by Aidan Gomez (CEO and a co-author of the Transformer paper), Ivan Zhang and Nick Frosst. The team started with three co-founders and three founding engineers, training its first models in early 2020 before there was any market for them. As of an October 2026 web search, Cohere has raised a $500 million Series D at a $5.5 billion valuation (July 2024), a $500 million round at $6.8 billion (August 2025) and a $100 million extension that took it to $7 billion (September 2025), and reported $240 million in annual recurring revenue for 2025. The one-line story: while the rest of the field chased chatbots and AGI, Cohere picked the boring, useful problems inside companies and stuck to them. Nick Frosst told the story on The Product Market Fit Show, recorded in 2024, shortly after the Series D.

What does Cohere do?

Cohere builds foundation models, the large language models that power chat, search and automation, and sells them to enterprises. It also builds the tools around those models that make them usable on a company's own data.

When the episode was recorded, Frosst described a handful of things Cohere focused on because of that enterprise lens:

Private deployment. Cohere will run its model on its own platform, on any major cloud, or on-premise, wherever the customer's data lives.
Retrieval augmented generation (RAG). The model answers questions from a company's own documents and gives citations, so people can check the answer.
Multilingual support across what Frosst called "all major business languages."
Tool use. The model can decide when to search, query a database or call another function, then use the result.
Embeddings and search, which help companies with huge piles of data find the relevant pieces.

As of an October 2026 web search, Cohere's product line includes the Command A family of models (including Command A Reasoning, Translate and Vision), the Aya multilingual models, and Cohere North, its workplace AI platform. It focuses on regulated industries such as finance, healthcare, manufacturing, energy and the public sector.

Frosst was blunt about what Cohere does not do:

"We just care about when we talk to a company who's using our model, is it useful for them? Is it doing the thing they need it to, to do? And that's what we obsess over and that's fairly unique within the foundational model companies." — Nick Frosst, Cohere

Key stat: as he shared on the show, Cohere had already seen customers use its technology to augment a workflow and get roughly a 50% increase in productivity.

Who founded Cohere?

Cohere was founded in 2019 by Aidan Gomez, Ivan Zhang and Nick Frosst, as of an October 2026 web search. Gomez, the CEO, was a research intern at Google Brain and one of the authors of "Attention Is All You Need," the 2017 paper that introduced the Transformer architecture behind today's language models. Frosst was one of Geoffrey Hinton's early hires at Google Brain in Toronto, where he worked as a machine learning researcher. Zhang had worked with Gomez at FOR.ai, an independent research group.

On the show, Frosst said the idea was Gomez's. Gomez saw that language models were a new kind of machine learning: for the first time, one general-purpose model trained on as much language as possible beat a narrow model built for a single task. If you wanted to pull numbers from a PDF or summarize a call transcript, the best tool was a general model. That meant someone needed to train one and make it available to businesses. Gomez then convinced Frosst and Zhang to quit their jobs.

"The market didn't really exist at that time. there was nobody selling access to large language models." — Nick Frosst, Cohere

They knew OpenAI would go after the same opportunity. They chose the enterprise anyway, because, as Frosst put it, that is where the technology adds the most value, both for internal work and for products companies build for their own customers.

Key stat: as he shared on the show, Cohere began with three co-founders and three founding engineers, and started working on language models in early 2020. By the time of the interview it had several hundred employees and offices around the world.

Frosst was CTO for a while but no longer held that title at the time of the interview. He said he fills gaps until Cohere hires someone better, then moves on.

How much has Cohere raised?

All figures below come from an October 2026 web search of press coverage (BetaKit, The Globe and Mail, CNBC, Bloomberg and company announcements). None of them were discussed on the show except the general fact of a large 2024 round.

Series C, 2023: $270 million.
Series D, July 2024: $500 million at a $5.5 billion valuation, led by PSP Investments, with NVIDIA, Fujitsu, Cisco, AMD Ventures and Export Development Canada. This is the round Pablo referred to in the episode as one of the largest in Canadian history.
August 2025: $500 million at a $6.8 billion valuation, led by Radical Ventures and Inovia Capital, with AMD Ventures, NVIDIA, PSP Investments, Salesforce Ventures and the Healthcare of Ontario Pension Plan.
September 2025: a $100 million extension that brought the valuation to $7 billion.
2025 revenue: Cohere reported $240 million in annual recurring revenue for 2025, beating its own $200 million forecast, according to an investor memo reported in February 2026.

Two bigger deals were still open as of October 2026:

Aleph Alpha merger. In April 2026 Cohere announced a deal to combine with German AI company Aleph Alpha, and in September 2026 the two signed a definitive agreement valuing the combined company at roughly $20 billion. It still needs regulatory approval and is expected to close later in 2026. The Schwarz Group committed about $600 million to Cohere's coming Series E as part of the deal.
New round. Bloomberg reported in September 2026 that Cohere was in advanced talks to raise $2 billion to $3 billion at a $20 billion valuation. As of October 2026 that round had not been confirmed as closed.

Key stat: as of September 2025, Cohere's last confirmed priced valuation was $7 billion, up from $5.5 billion at the time of the interview.

There is no MVP of a power plant

Frosst said the scrappy-MVP playbook does not fit a foundation model company. An untrained model is useless, and compute, data and talent are all expensive. He compared it to building a mine, an oil refinery or a power plant.

So Cohere's early goal was to prove to investors the team could do it. In the first months it trained on small numbers of GPUs spread across a data center, which was very slow. Everyone knew it would not scale.

"So the first few months we were training large language models in a way that we knew wouldn't scale because we didn't have access to compute, but it was good enough to say, Hey, look, we can do this." — Nick Frosst, Cohere

Frosst said that proof, showing the talent and the data were there and only the compute was missing, is what took Cohere from its seed round to its Series A.

ChatGPT was a usability breakthrough, not a capability one

Cohere had been building language models for about two and a half years when ChatGPT launched in late 2022. Frosst's take on that moment is one of the more useful parts of the episode for founders.

Before ChatGPT, base models could already do almost everything people now use them for. They were just hard to use: instead of asking for "a poem about a fish," you wrote the start of a fake document and let the model finish it.

Never miss a founder's PMF story

Subscribe to The PMF Show

What surprised him, and almost everyone outside OpenAI, was how little extra data it took to turn a base model into a chat model. A small amount of human feedback made the model far easier to use, and it generalized to requests it had never seen in training.

"That's a really interesting historical moment that what woke people up to the possibility of language models was an ease of use." — Nick Frosst, Cohere

Cohere's response was to build its own chat training data, but aimed at enterprise work rather than a consumer chatbot.

He is not worried about free open-weight models either. Cohere publishes its own weights for researchers. In his view, the value is the process that trains, deploys and keeps improving the models, not any one set of weights.

"So when I think about Cohere’s value, it's not in, you know, the weights of any particular model." — Nick Frosst, Cohere

Product-market fit came in pieces, and the boring use cases mattered most

Frosst pushed back on the idea of product-market fit as a single moment. He compared it to fitness in evolution: you can only see it looking back. Founders who convince themselves they have it early can mislead themselves.

Still, he named the moments that felt like it. The first was when Cohere released its API, people signed up, and they told the team what it was doing for them.

"And I got especially excited when the thing it was doing for them was boring. That's when I would get most excited." — Nick Frosst, Cohere

His examples were summarizing documents and pulling numbers from quarterly earnings reports into a table, tasks people used to do by hand. The feeling came back stronger with private, on-premise and cloud deployments, and a little with each new feature.

He is also skeptical that language models lead to AGI. He sees them as sequence models that predict the next word: extremely useful, but not intelligent the way people are. For Cohere, that view shapes the product.

"I still get very excited about it all the time, but I don't think it's a clear path to AGI. I don't think we're gonna be making digital gods anytime soon." — Nick Frosst, Cohere

Key lessons from Nick Frosst's playbook

1. Start with the problem, not the technology. His advice to AI founders: figure out what you are solving first, then ask whether an LLM helps. He expects "AI company" to become as meaningless a label as "internet company" or "app company." 2. Pick a lane the giants are not optimizing for. Cohere ignored consumer chatbots, AGI and public benchmarks, and built for privacy, citations, multiple languages and deployment wherever the data is. That focus is what set it apart. 3. Boring use cases are the strongest signal. The tasks nobody wants to do, summaries and data extraction, were the clearest proof of real value. 4. In capital-heavy markets, your first milestone is proof, not product. Cohere's early, slow training runs existed to show investors the team could do it with more money. 5. The "incumbents will crush you" fear is old. Frosst heard that Google would beat them to it when they started. He thinks incumbents are moving faster, but the myth is often pushed by the incumbents themselves.

"So you should know as you're going along that every challenge you're facing, that's the hardest challenge you've faced so far. And it's gonna feel all encompassing and then later you're gonna look back and you're gonna say how quaint, but the emotional impact will always be the same." — Nick Frosst, Cohere

His advice to his younger self: chill out. The stakes felt as high at six people as at several hundred, and it does not get easier.

For more on building AI companies that last, see how to tell if your AI startup is just a wrapper, building a moat for an AI startup, landing your first enterprise customer, and Neo Financial's run at Canada's big banks.

Listen to the full interview: He built Cohere into a $5.5B AI startup; How to Win in AI; & Why LLMs won't lead to AGI.

FAQ: Cohere

Q: What is Cohere? A: Cohere is a Toronto-based AI company that builds large language models and AI tools for enterprises, with a focus on private deployment, retrieval with citations, multilingual support and regulated industries. Its products include the Command A models, the Aya models and the Cohere North platform.

Q: Who is the CEO of Cohere? A: Aidan Gomez, a co-author of the Transformer paper, is Cohere's co-founder and CEO. He founded the company in 2019 with Ivan Zhang and Nick Frosst.

Q: How much funding has Cohere raised? A: Confirmed rounds include a $270 million Series C in 2023, a $500 million Series D at a $5.5 billion valuation in July 2024, a $500 million round at $6.8 billion in August 2025 and a $100 million extension at $7 billion in September 2025. As of October 2026, a reported $2 billion to $3 billion raise at a $20 billion valuation and a roughly $20 billion merger with Aleph Alpha were announced but not yet closed.

Q: Who are Cohere's investors? A: Investors include PSP Investments, Radical Ventures, Inovia Capital, NVIDIA, AMD Ventures, Cisco, Fujitsu, Salesforce Ventures, Export Development Canada and the Healthcare of Ontario Pension Plan. The Schwarz Group committed about $600 million to its coming Series E.

Q: How much revenue does Cohere make? A: Cohere reported $240 million in annual recurring revenue for 2025, ahead of its $200 million target, according to an investor memo reported in February 2026.

Sources: Listen to the Full Founder Story

Nick Frosst, Co-Founder of Cohere — former Google Brain researcher who helped build an enterprise-only foundation model company, on why ChatGPT was a usability breakthrough, why LLMs will not lead to AGI, and why boring use cases prove real value.